I argue that data becomes temporarily interesting by itself to some self-improving, but computationally limited, subjective observer once he learns to predict or compress the data in a better way, thus making it subjectively simpler and more beautiful. Curiosity is the desire to create or discover more non-random, non-arbitrary, regular data that is novel and surprising not in the traditional sense of Boltzmann and Shannon but in the sense that it allows for compression progress because its regularity was not yet known. This drive maximizes interestingness, the first derivative of subjective beauty or compressibility, that is, the steepness of the learning curve. It motivates exploring infants, pure mathematicians, composers, artists, dancers, comedians, yourself, and (since 1990) artificial systems.
Simple test: imagine making a hill-climbing algorithm that maximized the increase in proportional compressibility of a piece of music as you listened to it. So you’d feed in some random seed, and out would come the local maximum of “interestingness.” What would this local maximum look like? It would actually be quite unlike current music. A real piece of music might start by introducing simple themes and then elaborating on them. But this gives away all the compressibility at the start! Elaborating on a known theme is “boring.” Instead you should start with noisy data that reveals itself to be random deviations around a simple pattern!
Other problems include the unsubstantiated downplaying of enjoying something more than once and ignoring the relationship with emotions.
So although this sort of algorithm might enjoy something vaguely like music, it would prefer to listen to many variations of simple statistical patterns at the highest speed possible—I’m looking for something a bit more human.
I’m in shock right now at reading that abstract. The OPs question is one I’ve thought about a lot, and this matches my intuitions/hypotheses perfectly.
I think the next interesting step is to look into the compression process and figure out exactly what patterns/relationships/properties the compression algorithm uses and how. Once you have that, you have a fully predictive model of musical quality, and could use it to generate optimal music.
But of course, that’s all assuming that this model is actually right, and would entail quite a lot of work even if that’s the case. Still, very excited to see this idea suggested academically, I’ve never encountered it before.
How does this model explain/deal with the facts that people have different musical tastes and that (at least some) people have no taste in music at all, but rather appreciate music based on weather or not their peers like it.
The paper isn’t particularly long, if you haven’t read it.
It doesn’t attempt to explain music at a cultural level, only an individual one. You don’t need a theory of aesthetics to explain why people would decide to like whatever their peers do, there’s plenty of general psychology to cover that.
As for different musical tastes, the compression algorithm that the model is based around is subjective and adaptive. So mine can be different from yours (though there’s a fair amount that humans on the whole will tend to have in common), and yours can change over time (esp in response to new data).
In particular, if you’ve been exposed to a lot of e.g. reggae music, then your algorithm will likely be especially efficient at compressing reggae. So it will seem more accessible to you. If I’ve been exposed to only a little reggae, it will likely seem less accessible, but more interesting: the compression algorithm can detect the presence of order and structure, but still has work to do in uncovering and utilizing all the regularity that’s there. And if someone has never heard any music but classical, reggae could be incomprehensible gibberish to that person (read: they won’t like it), because it clashes with their existing model and expectations so drastically.
Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes
Simple test: imagine making a hill-climbing algorithm that maximized the increase in proportional compressibility of a piece of music as you listened to it. So you’d feed in some random seed, and out would come the local maximum of “interestingness.” What would this local maximum look like? It would actually be quite unlike current music. A real piece of music might start by introducing simple themes and then elaborating on them. But this gives away all the compressibility at the start! Elaborating on a known theme is “boring.” Instead you should start with noisy data that reveals itself to be random deviations around a simple pattern!
Other problems include the unsubstantiated downplaying of enjoying something more than once and ignoring the relationship with emotions.
So although this sort of algorithm might enjoy something vaguely like music, it would prefer to listen to many variations of simple statistical patterns at the highest speed possible—I’m looking for something a bit more human.
I’m in shock right now at reading that abstract. The OPs question is one I’ve thought about a lot, and this matches my intuitions/hypotheses perfectly.
I think the next interesting step is to look into the compression process and figure out exactly what patterns/relationships/properties the compression algorithm uses and how. Once you have that, you have a fully predictive model of musical quality, and could use it to generate optimal music.
But of course, that’s all assuming that this model is actually right, and would entail quite a lot of work even if that’s the case. Still, very excited to see this idea suggested academically, I’ve never encountered it before.
How does this model explain/deal with the facts that people have different musical tastes and that (at least some) people have no taste in music at all, but rather appreciate music based on weather or not their peers like it.
The paper isn’t particularly long, if you haven’t read it.
It doesn’t attempt to explain music at a cultural level, only an individual one. You don’t need a theory of aesthetics to explain why people would decide to like whatever their peers do, there’s plenty of general psychology to cover that.
As for different musical tastes, the compression algorithm that the model is based around is subjective and adaptive. So mine can be different from yours (though there’s a fair amount that humans on the whole will tend to have in common), and yours can change over time (esp in response to new data).
In particular, if you’ve been exposed to a lot of e.g. reggae music, then your algorithm will likely be especially efficient at compressing reggae. So it will seem more accessible to you. If I’ve been exposed to only a little reggae, it will likely seem less accessible, but more interesting: the compression algorithm can detect the presence of order and structure, but still has work to do in uncovering and utilizing all the regularity that’s there. And if someone has never heard any music but classical, reggae could be incomprehensible gibberish to that person (read: they won’t like it), because it clashes with their existing model and expectations so drastically.
From talking with a friend in psychology, related ideas have been around for a while.